This project uses a combination of Uvicorn, FastAPI (Python) and Docker to provide a reliable REST API for testing Microsoft's BitNet out locally!
It supports running the inference framework, running BitNet model benchmarks and calculating BitNet model perplexity values.
It's offers the same functionality as the Electron-BitNet project, however it does so through a REST API which devs/researchers can use to automate testing/benchmarking of 1-bit BitNet models!
Install Conda: https://anaconda.org/anaconda/conda
Initialize the python environment:
conda init
conda create -n bitnet python=3.9
conda activate bitnet
Install the Huggingface-CLI tool to download the models:
pip install -U "huggingface_hub[cli]"
Download one/many of the 1-bit models from Huggingface below:
huggingface-cli download 1bitLLM/bitnet_b1_58-large --local-dir app/models/bitnet_b1_58-large
huggingface-cli download 1bitLLM/bitnet_b1_58-3B --local-dir app/models/bitnet_b1_58-3B
huggingface-cli download HF1BitLLM/Llama3-8B-1.58-100B-tokens --local-dir app/models/Llama3-8B-1.58-100B-tokens
Build the docker image:
docker build -t fastapi_bitnet .
Run the docker image:
docker run -d --name ai_container -p 8080:8080 fastapi_bitnet
Once it's running navigate to http://127.0.0.1:8080/docs
Note:
If seeking to use this in production, make sure to extend the docker image with additional authentication security steps. In its current state it's intended for use locally.
Building the docker file image requires upwards of 40GB RAM for Llama3-8B-1.58-100B-tokens, if you have less than 64GB RAM you will probably run into issues.
The Dockerfile deletes the larger f32 files, so as to reduce the time to build the docker image file, you'll need to comment out the find /code/models/.... lines if you want the larger f32 files included.